Adsa Logo White Adsa Title White

Can the threshold on absolute fat residual improve the reliability of milk mid-infrared-predicted traits without using reference values?

L. Zhang




Can the threshold on absolute fat residual improve the reliability of milk mid-infrared-predicted traits without using reference values?
L. Zhang*1, C. F. Li2, F. Dehareng3, C. Grelet3, F. Colinet1, N. Gengler1, Y. Brostaux1, H. Soyeurt1. 1TERRA Teaching and Research Centre, University of Li�ge-Gembloux Agro-Bio Tech Gembloux, Belgium, 2Hebei Livestock Breeding Station Shijiazhuang, China, 3Valorisation of Agricultural Products Department, Walloon Agricultural Research Centre Gembloux, Belgium.

Many traits are currently predicted using milk mid-infrared (MIR) spectrometry. However, those predictions can be erroneous for many reasons such as wrong milk collection, bad storage, inappropriate spectrometers management or calibration equation. Comparing predictions and reference values allows detecting those problems. However, preparing and analyzing reference samples are expensive, time consuming and sometimes difficult, especially for indirect traits like CH4. Therefore, it is relevant to develop approaches based on predictions to detect abnormal values. This work attempted to study the interest of using a threshold of absolute fat residual to detect abnormal MIR predictions. A total of 346,818 milk MIR records were collected from Chinese Holstein cows and analyzed by Bentley FTS spectrometers. The fat content predicted by the manufacturer model being corrected for the bias and slope were assumed to be the control value. From standardized spectra, a second fat content was externally predicted. The working hypothesis is that this content as only based on spectral data will reflect problem/noise present in the spectral data. The absolute residual fat was calculated as the absolute difference between internal and external predictions of milk fat content. The improvement of reliability was assessed using the difference of root mean square error (RMSE) before and after applying a threshold of 0.3 g/dL of milk for the fat residual. RMSE differences for protein, monounsaturated (MFA), saturated (SFA), and unsaturated (UFA) fatty acids were 0.003, 0.023, 0.014, and 0.024 g/dL of milk, respectively. The correlation coefficient between the internal and external predicted phenotypes nearly stayed constant: 0.96, 0.95, 0.97 and 0.95 for protein, MFA, SFA, and UFA, respectively. The use of a threshold based on milk fat residual allowed detecting abnormal predictions but as those values were not so frequent, RMSE and correlation values were not deeply impacted. This cleaning is therefore of interest for dairy herd improvement organizations to ensure the quality of their MIR spectral database.

Keywords: milk components, MIR, prediction reliability.